Security1 publisher2 min readPublished
Six ACM authors put the open-source bottleneck on the maintainers judging AI output
Six authors writing for the ACM's Technology Policy Council say AI has made code and vulnerability reports cheap to produce while the people who decide what ships are volunteers at projects like Apache, whose 2024 revenue was $2,379,402.
The Watch · Security desk
What happened
- Six authors writing for the ACM's Technology Policy Council, among them Simson Garfinkel and Josiah Dykstra, say AI coding tools are making open source software harder to maintain and secure.
- The tools write code and find flaws quickly, but the maintainers who decide what enters a project's official release still have to judge that output themselves.
- Despite US and EU mandates, the vast majority of open source applications still do not generate or ship a software bill of materials.
Compiled by The WatchSomething wrong?How this is made
Why it matters
- exposure One flaw in a widely used component obliges every downstream application to get its own patch and re-release in sequence, so the operator at the end of the chain waits on the slowest maintainer in it.
- constraint An SBOM lists what is present. It does not say whether the project behind a component is funded, maintained or abandoned, so inventory work alone will not find a dependency nobody is reviewing.
- decision Firms that pay nothing for the projects in their build now pick between funding review capacity at those projects and accepting whatever queue their dependencies sit in.
- capability The models that produce fixes also produce attacks, and the authors expect no protection from limiting strong security models to trusted users. Gatekeeping is therefore not a defensive option.
Most projects take contributions from anyone. Commits come from a vetted group of trusted contributors who decide what enters the official codebase [3]. AI lowered the cost of producing a submission, and some of what arrives is poor [4]. Quality may improve as models improve, the authors allow, but a human still has to assess the added volume [5].
The same tools serve defense and attack. Google says its CodeMender agent contributed "72 security fixes to open source projects, including some as large as 4.5 million lines of code" [6]. Spread across April to October 2025, that is roughly ten fixes a month [2]. Models that find and patch flaws can also be used to build attacks [7], and the authors doubt that restricting top security models to trusted users will help, since general-purpose models keep getting stronger [8].
Open source sits inside phones, cars, cloud systems and AI platforms, and deferred maintenance on thinly funded projects leaves holes in what is built on top of them [16]. AI-found flaws push organizations to patch sooner [9]. An application assembled from many open source components needs each component patched and re-released, one after another [10]. A growing number of attackers are planting malicious packages and code in popular repositories [11].
The Linux Foundation brought in $292,217,236 in 2024 [12]. The Apache Software Foundation, which the authors cite as an example of a project running on volunteer labor and sponsorships, brought in $2,379,402 [13], or 0.81 percent of the Linux Foundation's total [1]. Most open source use is unpaid, the free-rider problem [14]. A Harvard Business School paper the authors cite estimates that firms would spend 3.5 times more on software without open source [15].
The paper does not quantify how much review load AI has added. Open source is governed project by project, so ecosystem-wide numbers on AI's effects are hard to collect [19]. It argues from how the contribution funnel works, without measured submission volumes.
The authors want projects to put more resources into documentation, packaging, fundraising and gathering requirements, because deciding what a project needs and getting its community to agree still falls to people [20]. They write that most institutions face perhaps their largest challenge in learning the governance, maintenance and security state of their own software, so they can spot critical dependencies before something fails [21].
What to watch
- Whether any project starts publishing counts of AI-generated submissions and reviewer hours. The paper offers no ecosystem-wide figures.
- Whether US or EU SBOM mandates get extended past component lists to maintenance and funding status.
- Whether CodeMender-style agents keep contributing at the April-to-October 2025 rate, and whether maintainers keep accepting them.